---
title: 'LingBot-VLA 2.0: Robotic Manipulation Model'
url: https://www.emergentmind.com/topics/lingbot-vla-2-0
type: topic
---

# LingBot-VLA 2.0: Robotic Manipulation Model

LingBot-VLA 2.0 is a vision-language-action foundation model line for real-world robotic manipulation. In early 2026 it was presented as a pragmatic VLA system built from a pretrained VLM backbone and an action expert, trained on about 20,000 hours of real-world data from 9 popular dual-arm robot configurations [2601.18692]. Subsequent work reframed LingBot-VLA 2.0 as an application-oriented upgrade with about 60,000 hours total, including 50,000 hours of robot trajectories spanning 20 robot configurations and 10,000 hours of egocentric human videos, together with expanded whole-body action support and future prediction as a proxy task [2607.06403]. Across these reports, the model family is defined by large-scale multimodal pretraining, conditioning on robot state and language, and an explicit focus on cross-task, cross-embodiment, and deployment-oriented performance.

## 1. Development and stated objectives

The early 2026 formulation presents LingBot-VLA 2.0 as a “pragmatic” VLA foundation model whose goals are broad generalization across tasks and platforms together with cost efficiency in data and GPU hours required for adaptation [2601.18692]. That version is built around about 20,000 hours of real-world data from 9 popular dual-arm robot configurations and is evaluated through a systematic assessment on 3 robotic platforms, each completing 100 tasks with 130 post-training episodes per task. The same work also emphasizes training efficiency, reporting an optimized codebase with a throughput of 261 samples per second per GPU under an 8-GPU training setup and a 1.5–2.8$\times$ speedup over existing VLA-oriented codebases.

Later work extends the scope substantially. It describes LingBot-VLA 2.0 as bridging the disparity between laboratory conditions and real-world applications through three functional domains: generalization across tasks and embodiments, expanded action space beyond standard dual-arm manipulation, and predictive dynamics modeling for improved temporal reasoning [2607.06403]. In that account, the pretraining corpus grows to about 60,000 hours total, the embodiment set expands to 20 robot configurations, and the control interface is broadened to include heads, waists, mobile bases, and dexterous hands in addition to arms and grippers.

Taken together, these descriptions define LingBot-VLA 2.0 less as a single immutable checkpoint than as a research program centered on scaling real-world robot data, improving spatial and temporal grounding, and making VLA training operationally efficient. A plausible implication is that the name denotes a family of closely related systems rather than a single frozen architecture.

## 2. Pretraining corpora, embodiment coverage, and annotation

The earlier LingBot-VLA 2.0 report uses a pretraining corpus of about 20,000 hours collected from 9 dual-arm robot configurations: AgiBot G1, AgileX, Galaxea R1Lite, Galaxea R1Pro, Realman Rs-02, Leju KUAVO 4 Pro, Qinglong, ARX Lift2, and Bimanual Franka [2601.18692]. These platforms vary in arm degrees of freedom, camera layouts, and action-space dimensionality. The paper describes a two-stage annotation pipeline in which multi-view videos are decomposed by human annotators into clips corresponding to atomic actions, static frames at the start and end are removed, and Qwen3-VL-235B-A22B is used to annotate task and subtask instructions.

The later application-oriented version expands the corpus composition and the preprocessing pipeline. It states that about 90,000 hours of robot data from 20 embodiments were first collected and then filtered down to 50,000 hours of high-quality robot trajectories [2607.06403]. The cleaning pipeline computes third-order finite differences (jerk) on action and state signals, checks velocity and acceleration Z-scores, discards episodes with overly noisy or unstable motion, removes episodes in which the robot is effectively static for more than 95% of the episode, verifies alignment between video and state signals by projecting the robot with the URDF and comparing it to the recorded video, and filters blur, severe occlusion, dropped frames, and multi-view misalignment. In parallel, an egocentric human-video pool of about 20,000 hours is reduced to approximately 10,000 hours after processing. For unlabeled egocentric videos, the pipeline uses egocentric SLAM to estimate camera intrinsics and extrinsics and then applies hand pose estimation to recover MANO hand parameters.

A central design in the later system is a unified 55-dimensional canonical vector for both state and action. The vector covers 14 dimensions of arm joint position, 14 dimensions of end-effector pose, 2 dimensions of gripper position, 12 dimensions of hand joint position, 4 dimensions of waist position, 2 dimensions of head position, 3 dimensions of mobility signal, and 4 reserved dimensions [2607.06403]. This shared representation allows heterogeneous embodiments to be mapped into a common interface with padding where certain degrees of freedom do not exist.

Language supervision also evolves. The later work states that pretraining videos are annotated automatically with Qwen3.6-27B, which segments each manipulation video into contiguous subtasks and generates instructions. The annotation scheme includes a closed vocabulary of 18 action labels, video-level task instructions, subtask-level descriptions, and object labels from an open vocabulary. The closed action vocabulary includes primitive actions such as move, pour, push, pull, rotate, open, close, attach, and detach, as well as auxiliary labels such as transit, idle, and other.

## 3. Core architecture and optimization

In its earlier formulation, LingBot-VLA 2.0 uses a pretrained VLM backbone, Qwen2.5-VL, together with an initialized “action expert” [2601.18692]. These modules are combined using a Mixture-of-Transformers architecture inspired by BAGEL. The paper states that vision-language processing and action processing are separated into distinct transformer pathways while being coupled through a shared self-attention mechanism for unified layer-wise modeling. The stated goal is to preserve rich semantic priors from the VLM while avoiding destructive interference between language/vision tokens and action tokens.

The input-output formulation is explicit. At time step $t$, the observation sequence is
$$
\mathbf{O}_t = [\mathbf{I}_t^1, \mathbf{I}_t^2, \mathbf{I}_t^3, \mathbf{T}_t, \mathbf{s}_t],
$$
where the three $\mathbf{I}_t$ terms are three-view robot images, $\mathbf{T}_t$ is the task instruction, and $\mathbf{s}_t$ is robot state or proprioception. The action output is a trajectory chunk
$$
\mathbf{A}_t = [\mathbf{a}_t, \mathbf{a}_{t+1}, \dots, \mathbf{a}_{t+T-1}],
$$
with action chunk length $T=50$ during pretraining, and the model learns the conditional distribution $p(\mathbf{A}_t \mid \mathbf{O}_t)$ [2601.18692].

Continuous action generation is trained with Flow Matching rather than discrete action classification. The action expert minimizes a flow-matching objective over noisy interpolants between Gaussian noise and the target action chunk. The same report states that the model follows $\pi_0$ in using blockwise causal attention over image-plus-language tokens, state tokens, and action tokens: within each block tokens attend bidirectionally, while across blocks attention is causal so that future action tokens do not leak into current observations.

The early work also foregrounds systems-level optimization. Distributed training uses FSDP, with special “shard groups” for the action expert modules inspired by HSDP from VeOmni in order to reduce communication overhead [2601.18692]. Reductions are performed in float32 for stability, storage and communication use bfloat16, and operator-level optimizations include FlexAttention for sparse multimodal attention and `torch.compile` for operator fusion and reduced kernel launch overhead.

## 4. Spatial grounding, whole-body control, and predictive dynamics

A major extension of LingBot-VLA 2.0 is the incorporation of explicit geometric and temporal supervision. In the earlier system, spatial improvement appears as a depth-based alignment variant in which visual tokens are distilled from LingBot-Depth [2601.18692]. Learnable queries over the three views are projected toward depth tokens, and the resulting distillation loss is intended to improve spatial reasoning for manipulation tasks involving geometry, occlusion, and precise placement.

The later system generalizes this idea into a future-prediction proxy task [2607.06403]. It appends two learnable queries, $[\mathbf{Q}_t,\mathbf{Q}_{t+T}]$, where $\mathbf{Q}_t$ represents the current observation and $\mathbf{Q}_{t+T}$ represents a future observation at horizon $T$. These queries are distilled from two teachers. The depth teacher, LingBot-Depth, provides geometric supervision by aligning the current and future queries with depth representations at the current and future frames. The video teacher, DINO-Video, provides temporally grounded semantic features. DINO-Video is described as initialized from DINOv3 and extended with block-wise causal temporal attention and 3D rotary positional embeddings, and it is trained on 5 million video clips across internet, egocentric, and robotic sources using video-adapted DINO/iBOT self-distillation.

This predictive-dynamics formulation is paired with a broader control interface. The later paper states that LingBot-VLA 2.0 supports whole-body degrees of freedom, specifically head motion, waist motion, mobile base movement, and dexterous hand action in addition to arm joints, end-effector pose actions, and gripper or hand actions [2607.06403]. The action expert in that system uses a token-level sparse MoE architecture with sigmoid-affinity routing, top-$K$ expert selection, and a bias-based, auxiliary-loss-free load-balancing mechanism. The paper’s rationale is to increase capacity without adding a load-balancing loss that could interfere with action learning.

The combined methodological picture is therefore one in which LingBot-VLA 2.0 moves from dual-arm continuous action prediction toward a more explicitly grounded and more embodiment-complete policy. This suggests that the system’s later versions are designed not only to map observations to actions, but also to internalize geometric cues and short-horizon future structure as part of control.

## 5. Benchmarking, scaling behavior, and reported performance

The earlier LingBot-VLA 2.0 report evaluates on a real-world benchmark built on GM-100 and on RoboTwin 2.0 simulation [2601.18692]. On the 3-platform real-world benchmark, average results are reported as follows: WALL-OSS at 4.05% success rate and 10.35% progress score, GR00T N1.6 at 7.59% and 15.99%, $\pi_{0.5}$ at 13.02% and 27.65%, LingBot-VLA 2.0 without depth at 15.74% and 33.69%, and LingBot-VLA 2.0 with depth at 17.30% and 35.41%. Per-platform results for the depth variant are 11.98% SR and 30.47% PS on Agibot G1, 18.93% SR and 40.36% PS on AgileX, and 20.98% SR and 35.40% PS on Galaxea R1Pro. On RoboTwin 2.0 simulation, the reported average success rates are 82.74% clean and 76.76% randomized for $\pi_{0.5}$, 86.50% and 85.34% for LingBot-VLA 2.0 without depth, and 88.56% and 86.68% for the depth variant.

The same paper reports several scaling and ablation findings. Pretraining-scale experiments from 3,000 to 20,000 hours show steady improvement in both success rate and progress score, with no saturation observed at 20,000 hours [2601.18692]. On 8 representative GM-100 tasks on Agibot G1, LingBot-VLA 2.0 with only 80 demonstrations per task is reported to outperform $\pi_{0.5}$ trained with the full 130-demo set. The paper also attributes improved robustness and spatial reasoning to the depth-aligned variant.

The later application-oriented work changes the evaluation setting from per-task post-training to a generalist setting in which a single policy is trained jointly on all tasks for each embodiment [2607.06403]. On Agilex Cobot Magic, it reports an overall average of 66.2 progress and 34.4 success, compared with 58.2 and 30.0 for LingBot-VLA-1.0, 59.1 and 32.2 for $\pi_{0.5}$, and 36.3 and 17.8 for GR00T N1.7. On Galaxea R1 Pro, the corresponding averages are 34.6 and 15.6 for LingBot-VLA 2.0, 32.7 and 15.6 for LingBot-VLA-1.0, 27.4 and 8.9 for $\pi_{0.5}$, and 16.4 and 5.6 for GR00T N1.7. The same study highlights object-grounding-heavy tasks such as “Retrieve keychain” on Agilex, which rises from 67.5 progress and 60.0 success in LingBot-VLA-1.0 to 100.0 and 100.0 in 2.0, and “Pick out toy bone,” which rises from 77.5 and 70.0 to 95.0 and 90.0 on Agilex and from 62.5 and 40.0 to 87.5 and 70.0 on Galaxea.

For long-horizon mobile manipulation, the later paper reports two tasks across two embodiments: “Sort objects into refrigerator” on Astribot S1 and “Stove cleaning” on Cobot Magic-ARX X5 [2607.06403]. Against $\pi_{0.5}$, LingBot-VLA 2.0 achieves 77.1 progress and 60.0 success versus 65.3 and 46.7 in-domain on Astribot S1, and 37.0 and 13.3 versus 30.3 and 6.7 out-of-distribution. On Cobot Magic-ARX X5, it reports 84.3 and 66.7 versus 79.9 and 60.0 in-domain, and 67.5 and 40.0 versus 62.5 and 33.3 out-of-distribution. These results are presented as evidence of cross-embodiment long-horizon mobile manipulation capability.

## 6. External assessments, failure modes, and adjacent model families

Although LingBot-VLA 2.0 is reported as a strong VLA baseline, external studies also identify concrete failure regimes. The VISTA paper treats LingBot-VLA as one of the major VLA baselines and argues that direct use of raw UMI data creates two mismatches: wrist-mounted fisheye observations are out of distribution for pretrained VLMs, and human-collected trajectories are often physically infeasible for the target robot [2606.04708]. In diagnostic experiments, LingBot-VLA’s performance is reported to drop by an average of 15.7 points across LIBERO and RoboTwin under a wrist-fisheye regime. On UMI-style simulation benchmarks, the reported LingBot-VLA scores are 0.499 on RoboTwin-UMI and 0.817 on LIBERO-UMI, averaging 0.658, while on real-robot evaluation across 20 UMI-collected tasks the average is 0.313. VISTA is presented as a successor-like training recipe that preserves scalable UMI data collection while adding visual alignment through UMI-VQA and physical alignment through embodiment-conditioned trajectory validation.

A common source of confusion is the relation between LingBot-VLA and LingBot-VA. The robustness study “Do World Action Models Generalize Better than VLAs?” explicitly classifies LingBot-VA as a world action model rather than a VLA [2603.22078]. In that taxonomy, LingBot-VA uses a Wan2.2-5B video-generation backbone, a flow-matching action head, policy training plus task-specific finetuning, autoregressive or interleaved generation, and absolute end-effector poses plus joints as its action representation. It achieves 74.2% total success on RoboTwin 2.0-Plus, with especially high scores under language, light, background, noise, and layout perturbations, but it is weaker on camera and robot-initial-state perturbations and substantially slower at inference than $\pi_{0.5}$.

A further adjacent line is LingBot-VA 2.0, which is presented as a native video-action foundation model rather than a VLA foundation model [2607.08639]. That system introduces a semantic visual-action tokenizer, causal pretraining from scratch, a sparse MoE backbone, and asynchronous closed-loop control with re-grounding, and it reports deployment rates up to 225 Hz. The existence of both LingBot-VLA and LingBot-VA branches indicates that the broader LingBot research program spans at least two model families: VLA policies centered on VLM backbones and action experts, and VA or WAM systems centered on native video-action dynamics.

The principal limitation that emerges across these external assessments is not a claim that LingBot-VLA 2.0 fails in general, but that its performance depends strongly on observation regime, embodiment alignment, and action-space realism. VISTA’s results suggest that raw domain shift in wrist-fisheye perception and unvalidated trajectories can materially degrade deployment performance, while the robustness study suggests that dynamic video priors can improve certain perturbation dimensions but at a substantial runtime cost in related LingBot-VA systems.

Source: https://www.emergentmind.com/topics/lingbot-vla-2-0